23/3/2017

GeoVisual Search

 

Search public satellite imagery for visually similar objects.

Ever wanted to quickly find all instances of wind turbines in Europe? Or orchards? Or football fields? Descartes Labs has just released a simple in-browser tool that allows you to do just that.

GeoVisual Search is essentially a search engine for satellite imagery. Using machine learning, it lets users visually query images for similar geographic features.

For example, let’s say I wanted to search the globe for satellite imagery of city beaches. To do so, I simply zoomed into the Australian coastline and selected a tile of beach in a city:

geo.PNG

Within seconds, the platform lit up with red dots to indicate similar tiles:

geo1.PNG

The platform is built on satellite imagery from Landsat, the National Agriculture Imagery Program (NAIP), and PlanetScope.

To catalogue similar tiles, the team used a deep learning algorithm that classified each tile based on its visual fingerprint – colours, edges, and other features unique to the tile. Then, once a tile is selected, the algorithm compares every other image to that fingerprint and returns the best matches.

The project is still in demo-mode and the developers are working on ways to increase accuracy and providing more inclusive counts. Next, they’re looking to focus on object detection at scale.

Explore GeoVisual Search here.

Comments